Augusto B. Corrêa

In AI planning, it is common to distinguish between planning domains and problem instances, where a "domain" is generally understood as a set of related problem instances. This distinction is important, for example, in generalised planning, which aims to find a single, general plan or policy that solves all instances of a given domain. In PDDL, domains and problem instances are clearly separated: the domain defines the types, predicate symbols, and action schemata, while the problem instance specifies the concrete set of (typed) objects, the initial state, and the goal condition. In this paper, we show that it is quite easy to define a PDDL domain such that any propositional planning problem instance, from any domain, becomes an instance of this (lifted) "universal" domain. We construct different formulations of the universal domain, and discuss their implications for the complexity of lifted domain-dependent or generalised planning.

Classical planning tasks are represented using a logical language. It is common to use a first-order representation, as this makes the description of the tasks more compact. Planners have long relied on translating these first-order representations into propositional ones in a process called grounding. Earlier work showed that grounding can improve the performance of classical planners. But this can also lead to exponentially larger encodings, and tasks with short plans become intractable simply because of the size of their representation.

In this thesis, we focus on lifted planning, which operates directly on the first-order representation. This bypasses the need for grounding, avoiding its computational overhead. We show how to build a lifted planner, called Powerlifted, that uses different techniques from database theory and logic programming to achieve state-of-the-art performance. First, we show how to perform a state-space search over the lifted representation. We address the lifted successor generation problem and show that this is equivalent to solving conjunctive queries. By exploiting the acyclicity of conjunctive queries, Powerlifted can generate successors efficiently in many domains. Second, we show how to use Datalog to compute delete-relaxation heuristics, such as the additive and the FF heuristics, directly over the lifted representation. Finally, we show how to translate other state-of-the-art techniques, such as preferred operators and width search, from the ground to the lifted setting.

We then show that some ideas from our work on lifted planning can be carried over to the ground setting. Using techniques implemented in Powerlifted, we show how to improve the grounding algorithms commonly used by classical planners. This is significant, as grounding is a crucial phase of most classical planners developed over the last two decades. Our new algorithm uses the grounding via solving paradigm from answer set programming, in which the grounding of a problem is decoupled into several logic programs, which must be solved individually to produce the propositional representation. In our experiments, our method outperforms other commonly used grounders from the literature.

Finally, we propose an extension to the classical planning formalism allowing for object creation within action effects. In practice, we show that Powerlifted can support this fragment almost effortlessly and that lifted planners, in general, seem to be a good fit for planning with object creation. Our experimental results show that object creation does not add overhead to Powerlifted while allowing for more expressive planning formalisms.

Lifted planning is usually defined as planning directly over a first-order representation. From the mid-1990s until the late 2010s, lifted planning was sidelined, as most of the state-of-the-art planners first ground the task and then solve it using a propositional representation. Moreover, it was unclear whether lifted planners could scale. But as planning problems become harder, they also become infeasible to ground. Recently, lifted planners came back into play, aiming at problems where grounding is a bottleneck. In this work, we survey recent advances in lifted planning. The main techniques rely either on state-space search or logic satisfiability. For lifted search-based planners, we show the direct connections to other areas of computer science, such as constraint satisfaction problems and databases. For lifted planners based on satisfiability, the advances in modeling are crucial to their scalability. We briefly describe the main planners available in the literature and their techniques.

Classical planning problems are defined using some specification language, such as PDDL. The domain expert defines action schemas, objects, the initial state, and the goal. One key aspect of PDDL is that the set of objects cannot be modified during plan execution. While this is fine in many domains, sometimes it makes modeling more complicated. This may impact the performance of planners, and it requires the domain expert to bound the number of required objects beforehand, which can be a challenge. We introduce an extension to the classical planning formalism, where action effects can create and remove objects. This problem is semi-decidable, but it becomes decidable if we can bound the number of objects in any given state, even though the state space is still infinite. On the practical side, we extend the Powerlifted planning system to support this PDDL extension. Our results show that this extension improves the performance of Powerlifted while supporting more natural PDDL models.

One key decision for heuristic search algorithms is how to balance exploration and exploitation. In classical planning, novelty search has come out as the most successful approach in this respect. The idea is to favor states that contain previously unseen facts when searching for a plan. This is done by maintaining a record of the tuples of facts observed in previous states. Then the novelty of a state is the size of the smallest previously unseen tuple. The most successful version of novelty search is best-first width search (BFWS), which combines novelty measures with heuristic estimates. An orthogonal approach to balance exploration-exploitation is to use several open-lists. These open-lists are ordered using different heuristic estimates, which diversify the information used in the search. The search algorithm then alternates between these open-lists, trying to exploit these different estimates. This is the approach used by LAMA, a classical planner that, a decade after its release, is still considered state-of-the-art in agile planning. In this paper, we study how to combine LAMA and BFWS. We show that simply adding the strongest open-list used in BFWS to LAMA harms performance. However, we show that combining only parts of each planner leads to a new state-of-the-art agile planner.

Classical planning tasks are commonly described in a first-order language. However, most classical planners translate tasks by grounding them and then rewriting them into a propositional language. In recent years, the grounding step has become a larger bottleneck. In this work, we study how to improve it. We build on top of the most common grounder for planning tasks which uses Datalog to find all reachable atoms and actions. Inspired by recent progress in lifted planning, database theory, and algorithmics, we develop a new method to ground these Datalog programs. Our algorithm can ground more instances than the baseline, and most tasks it cannot ground are out of reach from any ground planner.

In classical planning, the aim is to find a sequence of deterministic actions leading from the initial to a goal state. In this work, we consider the scenario where a party who knows the solution to a planning task, called the prover, wants to convince a second party, the verifier, that it has the solution without revealing any information about the solution itself. This is relevant in domains where privacy is important, for example when plans contain sensitive information or when the solution should not be revealed upfront. We achieve this by introducing a zero-knowledge protocol for plan existence. By restricting ourselves to tasks with polynomially-bounded plan length, we are able to construct a protocol that can be run efficiently by both prover and verifier. The resulting protocol does not rely on any reduction, has a constant number of rounds, and runs in time polynomial in the size of the task.

The goal of generalized planning is to find a solution that works for all tasks of a specific planning domain. Ideally, this solution is also efficient (i.e., polynomial) in all tasks. One possible approach is to learn such a solution from training examples and then prove that this generalizes for any given task. However, such proofs are usually pen-and-paper proofs written by a human. In our paper, we aim at automating these proofs so we can use a theorem prover to show that a solution generalizes for any task. Furthermore, we want to prove that this generalization works while still preserving efficiency. Our focus is on generalized potential heuristics encoding tiered measures of progress, which can be proven to lead to a find in a polynomial number of steps in all tasks of a domain. We show our ongoing work in this direction using the interactive theorem prover Isabelle/HOL. We illustrate the key aspects of our implementation using the Miconic domain and then discuss possible obstacles and challenges to fully automating this pipeline.

Debugging can be a painful task, especially when bugs only occur for large input files. We present Machetli, a tool to help with debugging in such situations. It takes a large input file and cuts away parts of it, while still provoking the bug. The resulting file is much smaller than the original, making the bug easier to find and fix. In our experience, Machetli was able to reduce planning tasks with thousands of actions to trivial tasks that could even be solved by hand. Machetli is an open-source project and it can be extended to other use cases such as debugging SAT solvers or LaTeX compilation bugs.

The goal of generalized planning is to find a solution that works for all tasks of a specific planning domain. Ideally, this solution is also efficient (i.e., polynomial) in all tasks. One possible approach is to learn such a solution from training examples and then prove that this generalizes for any given task. However, such proofs are usually pen-and-paper proofs written by a human. In our paper, we aim at automating these proofs so we can use a theorem prover to show that a solution generalizes for any task. Furthermore, we want to prove that this generalization works while still preserving efficiency. Our focus is on generalized potential heuristics encoding tiered measures of progress, which can be proven to lead to a find in a polynomial number of steps in all tasks of a domain. We show our ongoing work in this direction using the interactive theorem prover Isabelle/HOL. We illustrate the key aspects of our implementation using the Miconic domain and then discuss possible obstacles and challenges to fully automating this pipeline.

Potential functions are a general class of heuristics for classical planning. For satisficing planning, previous work suggested the use of descending and dead-end avoiding (DDA) potential heuristics, which solve planning tasks by backtrack-free search. In this work we study the complexity of devising DDA potential heuristics for classical planning tasks. We show that verifying or synthesizing DDA potential heuristics is PSPACE-complete, but suitable modifications of the DDA properties reduce the complexity of these problems to the first and second level of the polynomial hierarchy. We also discuss the implications of our results for other forms of heuristic synthesis in classical planning.

Lifted planners are useful to solve tasks that are too hard to ground. Still, computing informative lifted heuristics is difficult: directly adapting ground heuristics to the lifted setting is often too expensive, and extracting heuristics from the lifted representation can be uninformative. A natural alternative for lifted planners is to use width-based search. These algorithms are among the strongest for ground planning, even the variants that do not access the action model. In this work, we adapt best-first width search to the lifted setting and show that this yields state-of-the-art performance for hard-to-ground planning tasks.

Heuristics for lifted planning are not yet as informed as the best heuristics for ground planning. Recent work introduced the idea of using Datalog programs to compute the additive heuristic over lifted tasks. Based on this work, we show how to compute the more informed FF heuristic in a lifted manner. We extend the Datalog program with executable annotations that can also be used to define other delete-relaxation heuristics. In our experiments, we show that a planner using the lifted FF implementation produces state-of-the-art results for lifted planners. It also reduces the gap to state-of-the-art ground planners in domains where grounding is feasible.

Recent research in classical planning has shown the importance of search techniques that operate directly on the lifted representation of the problem, particularly in domains where the ground representation is prohibitively large. In this paper, we show how to compute the additive and maximum heuristics from the lifted representation of a problem. We do this by adapting well-known reachability analysis techniques based on a Datalog formulation of the delete relaxation of the problem. Our adaptation allows us to obtain not only the desired heuristic value, but also other useful heuristic information such as helpful actions. Our empirical evaluation shows that our lifted version of the additive heuristic is competitive with its ground counterpart on most of the standard international competition benchmarks, and significantly outperforms other state-of-the-art lifted heuristic methods in the literature.

Planning is an attractive approach to solving problems because of its generality and its ease of use. A domain expert often has knowledge that could improve the performance of such a solution but this knowledge may be vague or not directly actionable. For example, an expert could be aware that a certain property is important but not have a good way of using this knowledge to speed up the search for a solution. In this paper we evaluate concept languages as a possible input language for injecting expert domain knowledge into a planning system. We also explore mixed integer programming as a way to use this knowledge to improve search efficiency and to help the user find and refine useful domain knowledge.

The standard PDDL language for classical planning uses several first-order features, such as schematic actions. Yet, most classical planners ground this first-order representation into a propositional one as a preprocessing step. While this simplifies the design of other parts of the planner, in several benchmarks the grounding process causes an exponential blowup that puts otherwise solvable tasks out of reach of the planners. In this work, we take a step towards planning with lifted representations. We tackle the successor generation task, a key operation in forward-search planning, directly on the lifted representation using well-known techniques from database theory. We show how computing the variable substitutions that make an action schema applicable in a given state is essentially a query evaluation problem. Interestingly, a large number of the action schemas in the standard benchmarks are acyclic conjunctive queries, for which query evaluation is tractable. Our empirical results show that our approach is competitive with the standard (grounded) successor generation techniques in a few domains and outperforms them on benchmarks where grounding is challenging or infeasible.

Generalized planning aims at computing solutions that work for all instances of the same domain. In this paper, we show that several interesting planning domains possess compact generalized heuristics that can guide a greedy search in guaranteed polynomial time to the goal, and which work for any instance of the domain. These heuristics are weighted sums of state features that capture the number of objects satisfying a certain first-order logic property in any given state. These features have a meaningful interpretation and generalize naturally to the whole domain. Additionally, we present an approach based on mixed integer linear programming to compute such heuristics automatically from the observation of small training instances. We develop two variations of the approach that progressively refine the heuristic as new states are encountered. We illustrate the approach empirically on a number of standard domains, where we show that the generated heuristics will correctly generalize to all possible instances.

Potential heuristics are weighted functions over state features of a planning task. A recent study defines the complexity of a task as the minimum required feature complexity for a potential heuristic that makes a search backtrack-free. This gives an indication of how complex potential heuristics need to be to achieve good results in satisficing planning. However, these results do not directly transfer to optimal planning.

In this paper, we empirically study how complex potential heuristics must be to represent the perfect heuristic and how close to perfect heuristics can get with a limited number of features. We aim to identify the practical trade-offs between size, complexity and time for the quality of potential heuristics. Our results show that, even for simple planning tasks, finding perfect potential heuristics might be harder than expected.

In this work, we investigate the computation of optimal plans for delete-free tasks using relaxed decision diagrams. We introduce a new method to compute approximations of h+ in polynomial time on the width of the RDD and show that this approximation is a lower bound for the optimal solution. We analyze different strategies to construct decision diagrams and compare these approximations to h+.

For a given state space and admissible heuristic function h there is always a tie-breaking strategy for which A* expands the minimum number of states [Dechter and Pearl, 1985]. We say that these strategies have optimal expansion. Although such a strategy always exists it may depend on the instance, and we currently do not know a tie-breaker that always guarantees optimal expansion. In this paper, we study tie-breaking strategies for A*. We analyze common strategies from the literature and prove that they do not have optimal expansion. We propose a novel tie-breaking strategy using cost adaptation that has always optimal expansion. We experimentally analyze the performance of A* using several tie-breaking strategies on domains from the IPC and zero-cost domains. Our best strategy solves significantly more instances than the standard method in the literature and more than the previous state-of-the-art strategy. Our analysis improves the understanding of how to develop effective tie-breaking strategies and our results also improve the state-of-the-art of tie-breaking strategies for A*.